habibayman/audio-anomaly-cnn

Audio Anomaly Detection using PCEN + CNN Ensemble

This repository contains a deep learning system for multi-class audio anomaly detection across industrial machines. The model uses Per-Channel Energy Normalization (PCEN) features and a convolutional neural network (CNN) ensemble trained with group-aware cross-validation.

---

Overview

The system classifies audio recordings into six classes:

Class	Description
0	Machine 1 - Normal
1	Machine 1 - Abnormal
2	Machine 2 - Normal
3	Machine 2 - Abnormal
4	Machine 3 - Normal
5	Machine 3 - Abnormal

The design emphasizes robustness to varying recording conditions and avoids data leakage through group-aware validation.

---

Pipeline Architecture

1. Preprocessing

• Audio resampled to 16 kHz
• Converted to mono
• Silence trimming using a 30 dB threshold
• No amplitude normalization or pre-emphasis

Note: Amplitude normalization and pre-emphasis are intentionally excluded because they interfere with PCEN, which performs its own adaptive gain control.

---

2. Feature Extraction (PCEN)

• Mel spectrogram:
  • 128 Mel bands
  • Power = 2.0
• PCEN parameters:
  • Gain (α): 0.98
  • Bias (δ): 2
  • Compression (r): 0.5
  • Time constant (T): 0.400 s

Output processing:

• Min-max normalization to [0, 1]
• Resized to (128 × 128) using:
  • Center cropping if too long
  • Zero-padding if too short

---

3. Model Training

CNN architecture:

• Three convolutional blocks:
  • Conv → BatchNorm → ReLU → Pooling
  • Channels: 16 → 32 → 64
• Final layers:
  • Adaptive average pooling
  • Dropout (30%)
  • Fully connected layer (6 outputs)

Training configuration:

• Stratified Group K-Fold (3 folds)
• Class-weighted cross-entropy loss
• Adam optimizer:
  • Learning rate: 1e-4
  • Weight decay: 1e-4
• Early stopping (patience = 8)

Data augmentation:

• Mixup (p = 0.25, Beta(0.2, 0.2))
• Time and frequency masking
• Gaussian noise injection

---

4. Ensemble Inference

• Three fold-specific models are used
• Predictions are combined using soft voting:
  • Logits are averaged across models
  • Final prediction is obtained via argmax

This reduces variance and improves generalization compared to a single model.

---

Performance

Cross-validation (3 folds)

Fold	Accuracy
1	89.54%
2	91.24%
3	89.76%
Mean	90.18%

---

Test Set Results

• Accuracy: 92.55%

Class-wise observations:

• Strong performance on normal classes
• Abnormal classes show more variability
• Machine 2 abnormal class is the most challenging due to limited data and overlap

---

Inference

Default usage

Run the script directly:

python infer.py

This assumes the following structure:

├── infer.py
├── models/
│   ├── cnn_fold_2_acc_0.9124.pth
│   ├── data/
│   │   ├── 1.wav
│   │   ├── 2.wav
│   │   └── ...

Custom paths

You can override all paths via CLI arguments:

python infer.py \
  --data_dir /path/to/wavs \
  --model_path /path/to/model.pth \
  --results /path/to/results.txt \
  --times /path/to/time.txt

Input requirements

• .wav audio files only
• Files must be numerically named (e.g. 1.wav, 2.wav, ..., 10.wav)
• Any duration (automatically trimmed and resized)
• Audio is resampled internally to 16 kHz

Output

• results.txt: one predicted class per line
• time.txt: inference time (seconds) per file

---

Model Details

• Input shape: (1, 128, 128)
• Output: 6-class logits

Architecture:

• Conv2D → BatchNorm → ReLU → Pooling (×3)
• Adaptive average pooling
• Fully connected classification layer

---

Repository Structure

├── models/
│   ├── cnn_fold_*.pth
│   ├── results.txt
│   └── time.txt
├── infer.py
├── requirements.txt
└── README.md

---

Important Notes

• Preprocessing must match training exactly
• Do not apply amplitude normalization or pre-emphasis
• PCEN relies on raw signal dynamics

---

Limitations

• Lower performance on Machine 2 abnormal class
• Some inter-machine spectral overlap remains
• Performance depends on consistency of recording conditions

---

Future Work

• Improve minority class representation
• Explore alternative model architectures
• Enhance robustness to domain shifts
• Investigate transformer-based audio models